The democratic case for limiting an AI citizens’ assembly
My own attraction to citizens’ assemblies over other forms of deliberative engagement is rooted in going to a high school with a culture of Quaker consensus practice, in first-hand experiences of participatory budgeting, and more recently, in contemplating the French Citizens’ Convention for Climate. But even while cultivating this attraction, I have asked myself: just what role can and should citizens’ assemblies play in regulating and governing artificial intelligence? This question pushes beyond commitment to deliberative democracy in general, requiring a close look at where assemblies are particularly well suited. The answer depends in part on separating several functions that are sometimes grouped together under the general heading of AI governance.
There are several different types of task that assemblies might be asked to carry out. One is to articulate values—that is, to generate deliberative judgments about what society collectively values and how much risk it is prepared to bear in pursuing those values. This is a job for which assemblies are exceptionally well suited. It requires normative deliberation about collective purposes and acceptable levels of risk. Assemblies, if structured and utilized properly, can provide inclusive opportunities for people to engage in this difficult work, and exercise reflective judgment about what a society wants and ought to do. Nor does this task ordinarily require the assembly itself to diagnose a pressing technical incident, so the assembly’s ability to work slowly and thoughtfully during its deliberative period is an asset. The work of the French Citizens’ Convention for Climate, which translated a long-term emissions target into concrete policy proposals while deliberating about social justice and acceptable burdens, illustrates the capacity of assemblies to apply values to collective objectives.
Another is to conduct a diagnostic analysis of a technical incident. This requires swift assessment by those with specialized expertise. There are several impediments to conducting such an assessment by citizens’ assemblies. First, time is a constraint both because technical incidents typically require rapid analysis and resolution, and because assemblies tend to take place over extended periods. Second, the required diagnostic work is highly domain-specific and requires specialized knowledge that assemblies don’t usually possess. Third, while it is absolutely within the assembly’s capability to develop robust norms about safety and responsibility, sharing the information necessary for a technical diagnosis with a deliberative assembly might create unforeseen information hazards or other risks. All this does not mean that citizens are ignorant or incapable of grasping basic aspects of a technical incident, but it does suggest that the task at hand is not one that assemblies are well suited to perform.
Third, assemblies might be asked to choose an emergency response: that is, decide how to react to some event in order to mitigate its consequences. This task straddles a couple of potentially conflicting imperatives. On the one hand, the event often requires swift response. On the other, democratic legitimacy demands that public authority be exercised within publicly justified and democratically established constraints. One way of recognizing these two imperatives without leaving people to either face an emergency unassisted or delegate all the responses to officials is to have citizens’ assemblies help establish the permissible bounds of emergency action in advance: so that officials know what bounds they should observe as they respond swiftly to some event, and citizens are assured that whatever officials do stays within pre-established boundaries consistent with publicly established norms. Although in an emergency there may be little time for deliberation, an assembly convened in advance could help establish the parameters within which officials would be authorized to proceed.
Fourth, assemblies might be asked to review whether institutions acted legitimately. This involves retrospectively evaluating whether the institutions’ actions and decision-making processes met accepted social norms. This constitutes a natural site of work for deliberative assemblies, which can assess what counts as legitimate action in light of a set of commonly understood norms and procedures. Their involvement in this task complements their involvement in the first. While the first task sets norms for how action can be taken in the future, this one closes the loop by assessing whether institutions should be held accountable for how they acted in the past.
All this suggests that there is no need to reject citizens’ assemblies for AI governance altogether. Instead, they might be given a bounded but nonetheless useful role in the mix of responses envisaged. They could be used to articulate values upfront, and then to review whether institutions acted in accordance with those values. Such a division recognizes both democratic ideals and the functional demands of technical expertise, sensitive information, and emergency response.